Model Details
DeepSeek-V2.5
Quick answer: DeepSeek-V2.5 is DeepSeek's September 2024 MoE model — an intermediate release between V2 and V3. It scores 76.2% on Arena Hard and 74.7% on MATH. Available with open weights under the DeepSeek Model License, it was the state-of-the-art cost-efficient open model at launch, superseded by DeepSeek-V3 (December 2024).
At a Glance
Where DeepSeek-V2.5 leads
- 76.2% Arena Hard — strong general instruction following
- 74.7% MATH — good competition mathematics performance
- Open weights available on Hugging Face
- MoE architecture: cost-efficient inference relative to parameter count
- Predecessor that informed DeepSeek-V3's improvements
Where it lags
- Superseded by DeepSeek-V3 (90.2% MATH, 88.5% MMLU, MIT license) in all dimensions
- DeepSeek Model License (not MIT) — more restrictive than V3
- July 2024 knowledge cutoff
- Text-only
Best for: Historical comparisons; existing production integrations on V2.5; teams that have validated V2.5 for specific tasks and haven't migrated to V3.
What DeepSeek-V2.5 Is
DeepSeek-V2.5 is an intermediate release in DeepSeek's development roadmap, launched September 5, 2024. It combined the capabilities of DeepSeek-V2-Chat and DeepSeek-Coder-V2-Instruct into a single model, improving both general language capability and coding performance.
The model represents DeepSeek's state-of-the-art at Q3 2024, demonstrating strong Arena Hard (76.2%) and MATH (74.7%) scores competitive with leading models at the time. It was a key step in DeepSeek's trajectory toward V3, which launched December 2024 with dramatically improved scores across all benchmarks.
For new deployments, DeepSeek-V3 is strongly preferred: MIT license (vs DeepSeek Model License), significantly better benchmarks (90.2% MATH vs 74.7%), and similarly cost-efficient MoE architecture. DeepSeek-V2.5 is archived but remains available for teams with production integrations.
Specifications
| Field | Value |
|---|---|
| Organization | DeepSeek |
| Total parameters | ~236B (MoE) |
| Context window | 128,000 tokens |
| License | DeepSeek Model License |
| HuggingFace | deepseek-ai/DeepSeek-V2.5 |
| Release date | September 5, 2024 |
| Knowledge cutoff | July 2024 |
| Modality | Text only |
Public Benchmark Scores
| Benchmark | Score | Source | Date |
|---|---|---|---|
| Arena Hard | 76.2% | Benchgen evaluation | 2024-09 |
| MATH | 74.7% | Benchgen evaluation | 2024-09 |
DeepSeek-V2.5 vs Alternatives
| Model | MATH | Arena Hard | License |
|---|---|---|---|
| DeepSeek-V2.5 | 74.7% | 76.2% | DeepSeek ML |
| DeepSeek-V3 | 90.2% | — | MIT |
| Llama 3.1 405B Instruct | — | — | Llama 3.1 |
DeepSeek-V3 strictly dominates: 90.2% MATH vs 74.7%, MIT vs DeepSeek Model License, more recent knowledge cutoff. Migrate to V3 for all new deployments.
Related Models & Benchmarks
- Compare: DeepSeek-V3 · Llama 3.1 70B Instruct
- Benchmarks: MATH · Arena Hard
Frequently Asked Questions
What is DeepSeek-V2.5?
DeepSeek-V2.5 is DeepSeek's September 2024 MoE model scoring 76.2% Arena Hard and 74.7% MATH. It has been superseded by DeepSeek-V3, which achieves significantly better scores under a more permissive MIT license.Should I use DeepSeek-V2.5 or DeepSeek-V3?
DeepSeek-V3 for all new deployments: better benchmarks (90.2% MATH vs 74.7%), MIT license vs DeepSeek Model License, and more recent knowledge cutoff.Specs from DeepSeek's official V2.5 release (September 2024) and Benchgen evaluations. Last updated 2026-07-24.
Benchmark Leaderboards
This model isn’t on any benchmark leaderboard yet.